Abort postgres queries whose client died

docker/docker-compose.yml

@@ -7,6 +7,12 @@ services:
   postgres:
     image: postgres:15.0-alpine
     container_name: mpcautofill_postgres
+    # client_connection_check_interval (Postgres 14+): probes the client socket
+    # periodically DURING query execution and aborts the query if the client is
+    # gone, rather than computing to completion for a dead client (issue #462).
+    # Deliberately NOT statement_timeout - BULK-mode eligibility scans legitimately
+    # run for minutes-hours with a live client and must not be killed.
+    command: postgres -c client_connection_check_interval=30000
     ports:
       - "127.0.0.1:5432:5432"
     environment:

docs/infrastructure.md

@@ -107,6 +107,11 @@ this fixed and its follow-on hardening. Fixed by `eaece1fd` (#18,
   recreates the `django` container), also restart `nginx`** — see
   [[troubleshooting.md]] ("nginx 502s everything after a django container
   restart") for the mechanism and exact fix.
+- **`client_connection_check_interval=30000`** (postgres `command:`, issue
+  #462) makes a backend abort its query if the client socket died mid-query
+  instead of computing to completion for nobody. Rollout requires one
+  deliberate `postgres` container recreate (seconds of DB downtime) — django
+  and the worker reconnect on their own, no restart needed for either.
 
 ### Boot-time recovery